Improvements to the IBM Aurora 2 multi-condition system

Abstract

In this paper we describe some recent improvements to the performance of the Aurora 2 noisy digits speech recognition system for the matched training and test condition. The algorithms that we used pertain to discriminant acoustic modeling based on the Maximum Mutual Information (MMI) criterion, non-linear speaker/channel adaptation through probability distribution function matching. In addition, we revisited our last year's baseline system and improved its performance through cross-word context dependent modeling and Gaussian mixture components selection using the Bayesian Information Criterion (BIC). The aggregated result is 93.3% word accuracy for the multi-condition training data scenario.